floraachy/OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Cortex-A57: Fix clang compilation errors

Browse Source
This commit is contained in:
Ashwin Sekhar T K
2016-03-24 10:31:28 +05:30
parent 3b5ffb49d3
commit 278511ad2d
18 changed files with 3204 additions and 3204 deletions
Download Patch File Download Diff File Expand all files Collapse all files

204
kernel/arm64/dgemm_kernel_4x4.S
View File

@@ -161,150 +161,150 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ldp q0, q1, [pA]
add pA, pA, #32
fmul v16.2d, v0.2d, v8.2d[0]
fmul v29.2d, v1.2d, v11.2d[0]
fmul v16.2d, v0.2d, v8.d[0]
fmul v29.2d, v1.2d, v11.d[0]
ldp q2, q3, [ppA]
add ppA, ppA, #32
fmul v20.2d, v0.2d, v9.2d[0]
fmul v25.2d, v1.2d, v10.2d[0]
fmul v20.2d, v0.2d, v9.d[0]
fmul v25.2d, v1.2d, v10.d[0]
prfm PLDL1KEEP, [pA, #A_PRE_SIZE]
fmul v18.2d, v2.2d, v8.2d[0]
fmul v31.2d, v3.2d, v11.2d[0]
fmul v18.2d, v2.2d, v8.d[0]
fmul v31.2d, v3.2d, v11.d[0]
prfm PLDL1KEEP, [ppA, #A_PRE_SIZE]
fmul v22.2d, v2.2d, v9.2d[0]
fmul v27.2d, v3.2d, v10.2d[0]
fmul v22.2d, v2.2d, v9.d[0]
fmul v27.2d, v3.2d, v10.d[0]
ldp d12, d13, [pB]
add pB, pB, #16
fmul v24.2d, v0.2d, v10.2d[0]
fmul v21.2d, v1.2d, v9.2d[0]
fmul v24.2d, v0.2d, v10.d[0]
fmul v21.2d, v1.2d, v9.d[0]
ldp q4, q5, [pA] // for next round
add pA, pA, #32
fmul v26.2d, v2.2d, v10.2d[0]
fmul v23.2d, v3.2d, v9.2d[0]
fmul v26.2d, v2.2d, v10.d[0]
fmul v23.2d, v3.2d, v9.d[0]
ldp q6, q7, [ppA] // for next round
add ppA, ppA, #32
fmul v28.2d, v0.2d, v11.2d[0]
fmul v17.2d, v1.2d, v8.2d[0]
fmul v28.2d, v0.2d, v11.d[0]
fmul v17.2d, v1.2d, v8.d[0]
ldp d14, d15, [pB]
add pB, pB, #16
fmul v30.2d, v2.2d, v11.2d[0]
fmul v19.2d, v3.2d, v8.2d[0]
fmul v30.2d, v2.2d, v11.d[0]
fmul v19.2d, v3.2d, v8.d[0]
.endm
.macro KERNEL8x4_M2
fmla v16.2d, v4.2d, v12.2d[0]
fmla v29.2d, v5.2d, v15.2d[0]
fmla v16.2d, v4.2d, v12.d[0]
fmla v29.2d, v5.2d, v15.d[0]
ldp d8, d9, [pB]
add pB, pB, #16
fmla v18.2d, v6.2d, v12.2d[0]
fmla v31.2d, v7.2d, v15.2d[0]
fmla v18.2d, v6.2d, v12.d[0]
fmla v31.2d, v7.2d, v15.d[0]
ldp d10, d11, [pB]
add pB, pB, #16
fmla v20.2d, v4.2d, v13.2d[0]
fmla v25.2d, v5.2d, v14.2d[0]
fmla v20.2d, v4.2d, v13.d[0]
fmla v25.2d, v5.2d, v14.d[0]
prfm PLDL1KEEP, [pB, #B_PRE_SIZE]
fmla v22.2d, v6.2d, v13.2d[0]
fmla v27.2d, v7.2d, v14.2d[0]
fmla v24.2d, v4.2d, v14.2d[0]
fmla v21.2d, v5.2d, v13.2d[0]
fmla v22.2d, v6.2d, v13.d[0]
fmla v27.2d, v7.2d, v14.d[0]
fmla v24.2d, v4.2d, v14.d[0]
fmla v21.2d, v5.2d, v13.d[0]
ldp q0, q1, [pA]
add pA, pA, #32
fmla v26.2d, v6.2d, v14.2d[0]
fmla v23.2d, v7.2d, v13.2d[0]
fmla v28.2d, v4.2d, v15.2d[0]
fmla v17.2d, v5.2d, v12.2d[0]
fmla v26.2d, v6.2d, v14.d[0]
fmla v23.2d, v7.2d, v13.d[0]
fmla v28.2d, v4.2d, v15.d[0]
fmla v17.2d, v5.2d, v12.d[0]
ldp q2, q3, [ppA]
add ppA, ppA, #32
fmla v30.2d, v6.2d, v15.2d[0]
fmla v19.2d, v7.2d, v12.2d[0]
fmla v30.2d, v6.2d, v15.d[0]
fmla v19.2d, v7.2d, v12.d[0]
.endm
.macro KERNEL8x4_M1
fmla v16.2d, v0.2d, v8.2d[0]
fmla v29.2d, v1.2d, v11.2d[0]
fmla v16.2d, v0.2d, v8.d[0]
fmla v29.2d, v1.2d, v11.d[0]
ldp d12, d13, [pB]
add pB, pB, #16
fmla v18.2d, v2.2d, v8.2d[0]
fmla v31.2d, v3.2d, v11.2d[0]
fmla v18.2d, v2.2d, v8.d[0]
fmla v31.2d, v3.2d, v11.d[0]
ldp d14, d15, [pB]
add pB, pB, #16
fmla v20.2d, v0.2d, v9.2d[0]
fmla v25.2d, v1.2d, v10.2d[0]
fmla v20.2d, v0.2d, v9.d[0]
fmla v25.2d, v1.2d, v10.d[0]
prfm PLDL1KEEP, [pA, #A_PRE_SIZE]
fmla v22.2d, v2.2d, v9.2d[0]
fmla v27.2d, v3.2d, v10.2d[0]
fmla v22.2d, v2.2d, v9.d[0]
fmla v27.2d, v3.2d, v10.d[0]
prfm PLDL1KEEP, [ppA, #A_PRE_SIZE]
fmla v24.2d, v0.2d, v10.2d[0]
fmla v21.2d, v1.2d, v9.2d[0]
fmla v24.2d, v0.2d, v10.d[0]
fmla v21.2d, v1.2d, v9.d[0]
ldp q4, q5, [pA]
add pA, pA, #32
fmla v26.2d, v2.2d, v10.2d[0]
fmla v23.2d, v3.2d, v9.2d[0]
fmla v26.2d, v2.2d, v10.d[0]
fmla v23.2d, v3.2d, v9.d[0]
fmla v28.2d, v0.2d, v11.2d[0]
fmla v17.2d, v1.2d, v8.2d[0]
fmla v28.2d, v0.2d, v11.d[0]
fmla v17.2d, v1.2d, v8.d[0]
ldp q6, q7, [ppA]
add ppA, ppA, #32
fmla v30.2d, v2.2d, v11.2d[0]
fmla v19.2d, v3.2d, v8.2d[0]
fmla v30.2d, v2.2d, v11.d[0]
fmla v19.2d, v3.2d, v8.d[0]
.endm
.macro KERNEL8x4_E
fmla v16.2d, v4.2d, v12.2d[0]
fmla v25.2d, v5.2d, v14.2d[0]
fmla v18.2d, v6.2d, v12.2d[0]
fmla v27.2d, v7.2d, v14.2d[0]
fmla v16.2d, v4.2d, v12.d[0]
fmla v25.2d, v5.2d, v14.d[0]
fmla v18.2d, v6.2d, v12.d[0]
fmla v27.2d, v7.2d, v14.d[0]
fmla v20.2d, v4.2d, v13.2d[0]
fmla v29.2d, v5.2d, v15.2d[0]
fmla v22.2d, v6.2d, v13.2d[0]
fmla v31.2d, v7.2d, v15.2d[0]
fmla v20.2d, v4.2d, v13.d[0]
fmla v29.2d, v5.2d, v15.d[0]
fmla v22.2d, v6.2d, v13.d[0]
fmla v31.2d, v7.2d, v15.d[0]
fmla v24.2d, v4.2d, v14.2d[0]
fmla v17.2d, v5.2d, v12.2d[0]
fmla v26.2d, v6.2d, v14.2d[0]
fmla v19.2d, v7.2d, v12.2d[0]
fmla v24.2d, v4.2d, v14.d[0]
fmla v17.2d, v5.2d, v12.d[0]
fmla v26.2d, v6.2d, v14.d[0]
fmla v19.2d, v7.2d, v12.d[0]
fmla v28.2d, v4.2d, v15.2d[0]
fmla v21.2d, v5.2d, v13.2d[0]
fmla v30.2d, v6.2d, v15.2d[0]
fmla v23.2d, v7.2d, v13.2d[0]
fmla v28.2d, v4.2d, v15.d[0]
fmla v21.2d, v5.2d, v13.d[0]
fmla v30.2d, v6.2d, v15.d[0]
fmla v23.2d, v7.2d, v13.d[0]
.endm
.macro KERNEL8x4_SUB
@@ -315,28 +315,28 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ldp q0, q1, [pA]
add pA, pA, #32
fmla v16.2d, v0.2d, v8.2d[0]
fmla v29.2d, v1.2d, v11.2d[0]
fmla v20.2d, v0.2d, v9.2d[0]
fmla v25.2d, v1.2d, v10.2d[0]
fmla v16.2d, v0.2d, v8.d[0]
fmla v29.2d, v1.2d, v11.d[0]
fmla v20.2d, v0.2d, v9.d[0]
fmla v25.2d, v1.2d, v10.d[0]
ldp q2, q3, [ppA]
add ppA, ppA, #32
fmla v24.2d, v0.2d, v10.2d[0]
fmla v21.2d, v1.2d, v9.2d[0]
fmla v28.2d, v0.2d, v11.2d[0]
fmla v17.2d, v1.2d, v8.2d[0]
fmla v24.2d, v0.2d, v10.d[0]
fmla v21.2d, v1.2d, v9.d[0]
fmla v28.2d, v0.2d, v11.d[0]
fmla v17.2d, v1.2d, v8.d[0]
fmla v18.2d, v2.2d, v8.2d[0]
fmla v31.2d, v3.2d, v11.2d[0]
fmla v22.2d, v2.2d, v9.2d[0]
fmla v27.2d, v3.2d, v10.2d[0]
fmla v18.2d, v2.2d, v8.d[0]
fmla v31.2d, v3.2d, v11.d[0]
fmla v22.2d, v2.2d, v9.d[0]
fmla v27.2d, v3.2d, v10.d[0]
fmla v26.2d, v2.2d, v10.2d[0]
fmla v23.2d, v3.2d, v9.2d[0]
fmla v30.2d, v2.2d, v11.2d[0]
fmla v19.2d, v3.2d, v8.2d[0]
fmla v26.2d, v2.2d, v10.d[0]
fmla v23.2d, v3.2d, v9.d[0]
fmla v30.2d, v2.2d, v11.d[0]
fmla v19.2d, v3.2d, v8.d[0]
.endm
.macro SAVE8x4
@@ -422,17 +422,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ld1 {v0.2d, v1.2d}, [pA]
add pA, pA, #32
fmla v16.2d, v0.2d, v8.2d[0]
fmla v29.2d, v1.2d, v9.2d[1]
fmla v16.2d, v0.2d, v8.d[0]
fmla v29.2d, v1.2d, v9.d[1]
fmla v20.2d, v0.2d, v8.2d[1]
fmla v25.2d, v1.2d, v9.2d[0]
fmla v20.2d, v0.2d, v8.d[1]
fmla v25.2d, v1.2d, v9.d[0]
fmla v24.2d, v0.2d, v9.2d[0]
fmla v21.2d, v1.2d, v8.2d[1]
fmla v24.2d, v0.2d, v9.d[0]
fmla v21.2d, v1.2d, v8.d[1]
fmla v28.2d, v0.2d, v9.2d[1]
fmla v17.2d, v1.2d, v8.2d[0]
fmla v28.2d, v0.2d, v9.d[1]
fmla v17.2d, v1.2d, v8.d[0]
.endm
.macro SAVE4x4
@@ -482,10 +482,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ld1 {v0.2d}, [pA]
add pA, pA, #16
fmla v16.2d, v0.2d, v8.2d[0]
fmla v20.2d, v0.2d, v8.2d[1]
fmla v24.2d, v0.2d, v9.2d[0]
fmla v28.2d, v0.2d, v9.2d[1]
fmla v16.2d, v0.2d, v8.d[0]
fmla v20.2d, v0.2d, v8.d[1]
fmla v24.2d, v0.2d, v9.d[0]
fmla v28.2d, v0.2d, v9.d[1]
.endm
.macro SAVE2x4
@@ -572,10 +572,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ld1 {v0.2d, v1.2d}, [pA]
add pA, pA, #32
fmla v16.2d, v0.2d, v8.2d[0]
fmla v17.2d, v1.2d, v8.2d[0]
fmla v20.2d, v0.2d, v8.2d[1]
fmla v21.2d, v1.2d, v8.2d[1]
fmla v16.2d, v0.2d, v8.d[0]
fmla v17.2d, v1.2d, v8.d[0]
fmla v20.2d, v0.2d, v8.d[1]
fmla v21.2d, v1.2d, v8.d[1]
.endm
.macro SAVE4x2
@@ -610,8 +610,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ld1 {v0.2d}, [pA]
add pA, pA, #16
fmla v16.2d, v0.2d, v8.2d[0]
fmla v20.2d, v0.2d, v8.2d[1]
fmla v16.2d, v0.2d, v8.d[0]
fmla v20.2d, v0.2d, v8.d[1]
.endm
.macro SAVE2x2
@@ -643,7 +643,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ldr d0 , [pA]
add pA, pA, #8
fmla v16.2d, v8.2d, v0.2d[0]
fmla v16.2d, v8.2d, v0.d[0]
.endm
.macro SAVE1x2
@@ -674,8 +674,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ld1 {v0.2d, v1.2d}, [pA]
add pA , pA, #32
fmla v16.2d, v0.2d, v8.2d[0]
fmla v17.2d, v1.2d, v8.2d[0]
fmla v16.2d, v0.2d, v8.d[0]
fmla v17.2d, v1.2d, v8.d[0]
.endm
.macro SAVE4x1
@@ -705,7 +705,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
ld1 {v0.2d}, [pA]
add pA , pA, #16
fmla v16.2d, v0.2d, v8.2d[0]
fmla v16.2d, v0.2d, v8.d[0]
.endm
.macro SAVE2x1